Accurate brain age prediction with lightweight deep neural networks

Medical Image Analysis, Journal Year: 2020, Volume and Issue: 68, P. 101871 - 101871

Published: Oct. 19, 2020

Deep learning has huge potential for accurate disease prediction with neuroimaging data, but the performance is often limited by training-dataset size and computing memory requirements. To address this, we propose a deep convolutional neural network model, Simple Fully Convolutional Network (SFCN), of brain age using T1-weighted structural MRI data. Compared other popular architectures, SFCN fewer parameters, so more compatible small dataset 3D volume The architecture was combined several techniques boosting performance, including data augmentation, pre-training, model regularization, ensemble bias correction. We compared our overall approach widely-used machine models. It achieved state-of-the-art in UK Biobank (N = 14,503), mean absolute error (MAE) 2.14y 99.5% sex classification. also won (both parts of) 2019 Predictive Analysis Challenge prediction, involving 79 competing teams 2,638, MAE 2.90y). describe here details approach, its optimisation validation. Our can easily be generalised to tasks different image modalities, released on GitHub.

Language: Английский

---

Artificial Intelligence for Mental Health Care: Clinical Applications, Barriers, Facilitators, and Artificial Wisdom

Biological Psychiatry Cognitive Neuroscience and Neuroimaging, Journal Year: 2021, Volume and Issue: 6(9), P. 856 - 864

Published: Feb. 9, 2021

Language: Английский

---

Different scaling of linear models and deep learning in UKBiobank brain images versus machine-learning datasets

Nature Communications, Journal Year: 2020, Volume and Issue: 11(1)

Published: Aug. 25, 2020

Recently, deep learning has unlocked unprecedented success in various domains, especially using images, text, and speech. However, is only beneficial if the data have nonlinear relationships they are exploitable at available sample sizes. We systematically profiled performance of deep, kernel, linear models as a function size on UKBiobank brain images against established machine references. On MNIST Zalando Fashion, prediction accuracy consistently improves when escalating from to shallow-nonlinear models, further with deep-nonlinear models. In contrast, structural or functional scans, simple perform par more complex, highly parameterized age/sex across increasing sum, keep improving approaches ~10,000 subjects. Yet, nonlinearities for predicting common phenotypes typical scans remain largely inaccessible examined kernel methods.

Language: Английский

---

Optimising network modelling methods for fMRI

NeuroImage, Journal Year: 2020, Volume and Issue: 211, P. 116604 - 116604

Published: Feb. 13, 2020

A major goal of neuroimaging studies is to develop predictive models analyze the relationship between whole brain functional connectivity patterns and behavioural traits. However, there no single widely-accepted standard pipeline for analyzing connectivity. The common procedure designing based entails three main steps: parcellating brain, estimating interaction defined parcels, lastly, using these integrated associations parcels as features fed a classifier predicting non-imaging variables e.g., traits, demographics, emotional measures, etc. There are also additional considerations when correlation-based measures connectivity, resulting in supplementary utilising Riemannian geometry tangent space parameterization preserve connectivity; penalizing estimates with shrinkage approaches handle challenges related short time-series (and noisy) data; removing confounding from brain-behaviour data. These six steps contingent on each-other, optimise general framework one should ideally examine various methods simultaneously. In this paper, we investigated strengths short-comings, both independently jointly, following measures: parcellation techniques four kinds (categorized further depending upon number parcels), five decision staying ambient matrices or space, choice applying estimators, alternative handling confounds finally novel classifiers/predictors. For performance evaluation, have selected two largest datasets, UK Biobank Human Connectome Project resting state fMRI data, run more than 9000 different variants total ∼14000 individuals determine optimum pipeline. independent validation, some best-performing ABIDE ACPI datasets (∼1000 subjects) evaluate generalisability proposed network modelling methods.

Language: Английский

---

Individual-Specific Areal-Level Parcellations Improve Functional Connectivity Prediction of Behavior

Cerebral Cortex, Journal Year: 2021, Volume and Issue: 31(10), P. 4477 - 4500

Published: March 31, 2021

Resting-state functional magnetic resonance imaging (rs-fMRI) allows estimation of individual-specific cortical parcellations. We have previously developed a multi-session hierarchical Bayesian model (MS-HBM) for estimating high-quality network-level Here, we extend the to estimate areal-level While parcellations comprise spatially distributed networks spanning cortex, consensus is that parcels should be localized, is, not span multiple lobes. There disagreement about whether strictly contiguous or noncontiguous components; therefore, considered three MS-HBM variants these range possibilities. Individual-specific estimated using 10 min data generalized better than other approaches 150 out-of-sample rs-fMRI and task-fMRI from same individuals. connectivity derived also achieved best behavioral prediction performance. Among variants, exhibited resting-state homogeneity most uniform within-parcel task activation. In terms prediction, gradient-infused was numerically best, but differences among were statistically significant. Overall, results suggest MS-HBMs can capture behaviorally meaningful parcellation features beyond group-level Multi-resolution trained models are publicly available (https://github.com/ThomasYeoLab/CBIG/tree/master/stable_projects/brain_parcellation/Kong2022_ArealMSHBM).

Language: Английский

---

Machine-learning-based diagnostics of EEG pathology

NeuroImage, Journal Year: 2020, Volume and Issue: 220, P. 117021 - 117021

Published: June 10, 2020

Machine learning (ML) methods have the potential to automate clinical EEG analysis. They can be categorized into feature-based (with handcrafted features), and end-to-end approaches learned features). Previous studies on pathology decoding typically analyzed a limited number of features, decoders, or both. For I) more elaborate analysis, II) in-depth comparisons both approaches, here we first develop comprehensive framework, then compare this framework state-of-the-art methods. To aim, apply proposed deep neural networks including an EEG-optimized temporal convolutional network (TCN) task pathological versus non-pathological classification. robust comparison, chose Temple University Hospital (TUH) Abnormal Corpus (v2.0.0), which contains approximately 3000 recordings. The results demonstrate that achieve accuracies same level as networks. We find across in astonishingly narrow range from 81--86\%. Moreover, visualizations analyses indicated used similar aspects data, e.g., delta theta band power at electrode locations. argue current binary decoders could saturate near 90\% due imperfect inter-rater agreement labels, such are already clinically useful, areas where experts rare. make available open source thus offer new tool for machine research.

Language: Английский

---

Deep learning encodes robust discriminative neuroimaging representations to outperform standard machine learning

Nature Communications, Journal Year: 2021, Volume and Issue: 12(1)

Published: Jan. 13, 2021

Recent critical commentaries unfavorably compare deep learning (DL) with standard machine (SML) approaches for brain imaging data analysis. However, their conclusions are often based on pre-engineered features depriving DL of its main advantage - representation learning. We conduct a large-scale systematic comparison profiled in multiple classification and regression tasks structural MRI images show the importance DL. Results that if trained following prevalent practices, methods have potential to scale particularly well substantially improve compared SML methods, while also presenting lower asymptotic complexity relative computational time, despite being more complex. demonstrate embeddings span comprehensible task-specific projection spectra consistently localizes task-discriminative biomarkers. Our findings highlight presence nonlinearities neuroimaging can exploit generate superior representations characterizing human brain.

Language: Английский

---

Shared and unique brain network features predict cognitive, personality, and mental health scores in the ABCD study

Nature Communications, Journal Year: 2022, Volume and Issue: 13(1)

Published: April 25, 2022

Abstract How individual differences in brain network organization track behavioral variability is a fundamental question systems neuroscience. Recent work suggests that resting-state and task-state functional connectivity can predict specific traits at the level. However, most studies focus on single traits, thus not capturing broader relationships across behaviors. In large sample of 1858 typically developing children from Adolescent Brain Cognitive Development (ABCD) study, we show predictive features are distinct domains cognitive performance, personality scores mental health assessments. On other hand, within each domain predicted by similar features. Predictive models generalize to measures same domain. Although tasks known modulate connectome, between resting task states. Overall, our findings reveal shared account for variation broad behavior childhood.

Language: Английский

---

Predicting Alzheimer's disease progression using deep recurrent neural networks

NeuroImage, Journal Year: 2020, Volume and Issue: 222, P. 117203 - 117203

Published: Aug. 4, 2020

Early identification of individuals at risk developing Alzheimer's disease (AD) dementia is important for disease-modifying therapies. In this study, given multimodal AD markers and clinical diagnosis an individual from one or more timepoints, we seek to predict the diagnosis, cognition ventricular volume every month (indefinitely) into future. We proposed applied a minimal recurrent neural network (minimalRNN) model data The Disease Prediction Of Longitudinal Evolution (TADPOLE) challenge, comprising longitudinal 1677 participants (Marinescu et al., 2018) Neuroimaging Initiative (ADNI). compared performance minimalRNN four baseline algorithms up 6 years Most previous work on predicting progression ignore issue missing data, which prevalent in data. Here, explored three different strategies handle Two treated as "preprocessing" issue, by imputing using timepoint ("forward filling") linear interpolation ("linear filling). third strategy utilized itself fill both during training testing ("model filling"). Our analyses suggest that with "model filling" favorably algorithms, including support vector machine/regression, state space (LSS) model, long short-term memory (LSTM) model. Importantly, although procedure found trained exhibited similar performance, when only 1 input 4 suggesting our approach might well just cross-sectional An earlier version was ranked 5th (out 53 entries) TADPOLE challenge 2019. current 2nd out 63 entries June 3rd, 2020.

Language: Английский

---

Cross-ethnicity/race generalization failure of behavioral prediction from resting-state functional connectivity

Science Advances, Journal Year: 2022, Volume and Issue: 8(11)

Published: March 16, 2022

Algorithmic biases that favor majority populations pose a key challenge to the application of machine learning for precision medicine. Here, we assessed such bias in prediction models behavioral phenotypes from brain functional magnetic resonance imaging. We examined using two independent datasets (preadolescent versus adult) mixed ethnic/racial composition. When predictive were trained on data dominated by white Americans (WA), out-of-sample errors generally higher African (AA) than WA. This toward WA corresponds more WA-like brain-behavior association patterns learned models. AA only, compared training only or an equal number and participants, accuracy improved but stayed below Overall, results point need caution further research regarding current minority populations.

Language: Английский

---